JP4222066B2 - Face matching device - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a face collation apparatus that collates and discriminates a registrant using face data such as eyes, nose, mouth, contour, facial expression, glasses, etc., and more specifically, while maintaining collation discrimination accuracy, The present invention relates to a face collation apparatus that manages by adding and deleting registration data as a database.
[0002]
[Prior art]
In recent years, this type of face collation apparatus has registered registrant-specific face data for collation discrimination when managing the registration of face data. However, in order to improve recognition accuracy even if the registrant is added, When a plurality of face data is registered, the data is increased.
[0003]
For this reason, when registering face data, instead of registering face images as they are, it is possible to reduce the amount of registered data by extracting and registering only characteristic data. Is also deleted, and the collation result is adversely affected. As a result, the performance of the face collation apparatus is deteriorated.
[0004]
For example, in the case of a person recognition apparatus provided with a means for updating registered data shown in Patent Document 1, a determination is made as to whether updating is possible at the same time as authentication, and if it is determined that updating is possible, based on input face data and current registration data By generating new update registration data and performing update registration, the accuracy of the recognition rate is kept high while suppressing an increase in the amount of registration data.
[0005]
However, even in this case, when the collation degree value of the input face data is low, it is described that update registration is performed by adding the face data input as it is to the current registration data. As a result, the registration data The amount was increasing. This limits the number of registrants that can be newly registered. In particular, when the registration data increases, it cannot be applied to a system having a small storage capacity, and the amount of registration data needs to be reduced. In addition, an increase in storage capacity can be considered, but in this case, the cost increases.
[0006]
Further, as shown in Patent Document 2, in a dictionary (registered data) provided for collation recognition, a partial dictionary (feature data) that includes a large amount of mutual information remains, and a partial dictionary that includes a small amount of mutual information A partial dictionary deletion method of deleting is known. This is a method of leaving the partial dictionaries as the number of similar partial dictionaries decreases in relation to the partial dictionaries in the dictionary, and deleting the partial dictionaries as the number of similar partial dictionaries increases.
[0007]
However, in this case, since the configuration is such that, among the partial dictionaries of registered data, the number of similar partial dictionaries is uniformly deleted, the amount of data decreases, but when applied to a face matching device, Since the human face has changed due to the change, it is deleted from the face data close to the present which is usually similar, and the past face data which is not similar is left. For this reason, the recognition performance is inevitably lowered.
[0008]
[Patent Document 1]
JP-A-11-175718
[Patent Document 2]
JP 2000-90271 A.
[0009]
[Problems to be solved by the invention]
As described above, the above two patent documents can maintain the collation discrimination accuracy, but the data amount increases or the collation discrimination accuracy decreases although the data amount decreases, but the data amount decreases and the collation discrimination accuracy. Was not satisfied with both maintenance. Moreover, even if the above-mentioned two patent documents and other similar publications including these are combined, the means for reducing the data amount while maintaining the collation discrimination accuracy is not suggested, so that the problem is solved. I can't.
[0010]
Therefore, in the present invention, when managing registration data, considering that the registration data is human face data that changes slightly due to secular change, the time when the registration data has been registered and the source used for discrimination It is an object of the present invention to provide a face collation apparatus that can delete registered data based on the elapsed time since.
[0011]
[Means for Solving the Problems]
The present invention relates to a storage means for storing face data composed of a plurality of feature amount data and registration time in association with all registrants, giving importance to the feature amount data, and for collation. Determining means for comparing the inputted face data with the face data stored in the storage means to obtain a collation degree value, and discriminating whether the person is a registrant based on the obtained collation degree value; and the storage means When the face data is newly registered, the face data is registered with the time at the time of registration, and when determining whether or not the determination means is a registrant, it is stored as a basis for determination. Time registration means for re-registering the face data with the time when it was determined, elapsed time calculation means for calculating the elapsed time up to that based on the latest time assigned to the face data, and the elapsed time But predetermined A face collation apparatus comprising: feature data deletion means for extracting face data that is longer than a predetermined time and deleting the feature data with low importance from the face data stored in the storage means. And
[0012]
Here, the feature data is extracted from the face image of the person including the facial expression and the presence or absence of eyeglasses in addition to the shape, position and size of the registrant's eyes, nose, mouth, cheek, contour, eyebrows, etc. It is data. The face data is a collection of the feature amount data.
[0013]
The importance is assigned to each feature quantity data, and is data representing how important the feature quantity data of a certain registrant is compared to the feature quantity data of other registrants. When the feature amount data of a registrant is close to the feature amount data of another registrant, the importance level is set low, and when the feature amount data is far from the feature value data of other registrants, the importance level is set high. As a result, the feature amount data set with a high degree of importance is feature amount data that is not so much for other registrants. Therefore, it can be said that the feature amount data can easily identify the registrant at the time of collation determination.
[0014]
The feature data deletion means is for deleting registered data having a lower importance from face data stored in the storage unit, for example, deleting one or more feature data having the lowest importance. The control unit can control the registration data.
[0015]
According to the present invention, when newly registering face data, the registration time is given in accordance with the face data, so that the time elapsed since the face data was registered can be obtained. Also, when determining whether or not the determination means is a registrant, the face data is used for determination because the stored face data used for determination is re-registered with the time when the determination was made. You can also see how much time has passed since then. The face data whose elapsed time is equal to or longer than a certain time is data that is out of date or data that is not used for collation discrimination, and is data that does not affect the collation discrimination accuracy even if it is deleted. The feature amount data with low importance is data that is not important for specifying a registrant at the time of collation determination because it is close to the feature amount data of other registrants. Therefore, by extracting face data whose elapsed time is more than a certain time and deleting in order from feature data with lower importance, there is no effect on collation discrimination accuracy, and a free data space is secured in the registered capacity. You can renew your registration.
[0016]
As a result, since the registered data capacity does not increase, it is possible to perform appropriate data management such that data search does not take time even when face matching is performed, and matching can be efficiently performed in a short time. Therefore, the present invention can be applied even when the data capacity is a small-scale system, and the cost is not increased.
[0017]
When human face data that changes slightly due to secular changes is taken into account, face data whose elapsed time from when it was newly registered or elapsed time from when it was used for discrimination is more than a certain time It is appropriate to extract and delete in ascending order of importance. As a result, registration data suitable for collation determination remains, and only important registration data can be effectively managed to identify a registrant at the time of collation determination. For this reason, the collation discrimination accuracy can be maintained without increasing the registration data.
[0018]
As another embodiment of the present invention, it is possible to provide a time registering unit that gives the time when the determination is made to the original face data when it can be determined that the determination unit is a registrant.
[0019]
When it is possible to collate and determine that the person is a registrant at the time of collation, it can be said that the registered data used for the discrimination is valid data with high reliability. Therefore, this data is re-registered with the time when it was determined. Even in the registration data used as the basis for discrimination, the time when the discrimination was not performed is not given to the registration data that cannot be verified as a registrant at the time of verification. For example, when a certain non-registrant performs a collation determination, since it is a non-registered person, it cannot be collated with a registrant, and the time when the determination is made is not given to the data used for the determination. In this way, it is ensured that only valid data with high reliability remains, and unnecessary data can be deleted with an elapsed time of a certain time or more.
[0020]
As another embodiment of the present invention, the importance is assigned to each feature quantity data, and based on an average value of collation degree values obtained for each feature quantity data by comparing the feature quantity data between the registrants. Further, it is also possible to obtain the distance degree data in which the average values of the matching degree values obtained for each feature amount data are separated by comparing the feature amount data of a certain registrant and other registrants.
[0021]
In this way, the importance level, which is data indicating how important the feature value data of a registrant is compared with the feature value data of other registrants, is expressed as the collation degree value between registrants. Based on the average value, it is easy to calculate and can be easily configured by using the separation degree data in which the matching degree value between a certain registrant and other registrants is separated.
[0022]
DETAILED DESCRIPTION OF THE INVENTION
An embodiment of the present invention will be described below with reference to the drawings.
The drawing shows a face collation device that captures and collates a human face. In FIG. 1, the face collation device 11 connects a camera 12 and a result display 13 via a connection cable 14 or wirelessly for collation management. It is connected to the control device 15 and is installed, for example, as a means for authenticating a registrant who uses the entrance / exit.
[0023]
The camera 12 described above is set so that a human face is photographed using a CCD camera or the like, and a face image is acquired when the person approaches the front surface of the camera 12. For example, photographing is started when it is detected that a person is approaching by a human detection sensor (not shown) such as a reflective infrared sensor. In addition to this, when the registrant himself presses down the operation button or operation key for starting collation, it is possible to determine that the registrant has requested collation and start photographing by the camera 12.
[0024]
The face image acquired in this way is analyzed by the control device 15 and collation is started. At this time, the control device 15 pre-registers and stores registration data of face images of all registrants. Are compared with the face image data actually obtained from the camera 12 at the time of collation, and the collation result of the determined face image match / mismatch is displayed and output on the result display 13.
[0025]
The above-described control device 15 can be configured by, for example, a personal computer, and this control device 15 may be installed near the face image data acquisition position where the camera 12 is installed and managed for collation. You may make it manage with a center apparatus. Furthermore, a plurality of cameras 12 can be connected to one control system for collation management.
[0026]
FIG. 2 shows a control circuit block diagram of the face collation apparatus 11. The control device 15 that controls the face collation apparatus 11 transmits a raw image obtained from the face image data from the camera 12 via an I / F (interface) unit 21. Are temporarily stored in the image memory 23 of the control unit 22.
[0027]
After the face position is detected by matching based on a feature model such as eyes, nose, mouth, etc. from the accumulated raw image, the face area captured by the face extraction unit 24 is extracted.
[0028]
After extracting the face area, the feature amount extraction unit 25 extracts feature amount data obtained as a collection of unique points cut out from the face image. The feature amount data is extracted by using a statistical method such as principal component analysis to extract a difference from the average face, or extracted from a grayscale image of eyes, nose, mouth, etc. by template matching.
[0029]
The extracted feature amount data and the feature amount data registered in advance for collation determination and stored in the storage unit 30 are compared and collated by the person specifying unit 26 to obtain a collation degree value. It is determined whether or not the user is a registered person based on the value, and the determination result is displayed on the result display 13.
[0030]
By the way, when managing the registration data in the storage unit 30, the feature data deletion unit 27 is provided that deletes the feature data in ascending order of importance in consideration of the validity of the registration data from the elapsed time of the registration data. Yes.
[0031]
The feature amount data deletion unit 27 first calculates the elapsed time so far based on the latest time given to the face data by the elapsed time calculation unit 28, and the calculated elapsed time is a predetermined elapsed time. Face data with a low collation value that is longer than the time is extracted, and feature quantity data having the lowest importance is deleted from the face data stored in the storage unit 30.
[0032]
At the time of new registration of face data in the storage unit 30, the face data is registered with the time of registration timed by the timer T. Further, when determining whether or not the user is a registrant, the person identifying unit 26 re-registers the stored face data used for the determination with the time when the determination is made.
[0033]
In addition, at the time of new registration in the storage unit 30, the face data is associated with all the registrants, and importance (to be described later) (see FIG. 5) is assigned to the feature amount and stored.
[0034]
In addition, the input unit 29 allows an ID code input operation or an ID card input operation operated at the start of registration, and data for authenticating the registrant is input to the control unit 22.
[0035]
A method for obtaining the above-described importance will be described with reference to FIGS. FIG. 3 is a diagram when the face data of Mr. X is newly registered in the registered face data of Mr. A, Mr. B, Mr. C, etc. stored in the storage unit 30. In this case, the importance is assigned to each feature amount data of the face data of Mr. X to be newly registered and stored in the storage unit 30. First, based on the face image of Mr. X taken by the camera 12, the feature amount extraction unit 25 uses the feature amount extraction unit 25 based on the shape, position, and size of eyes, nose, mouth,. Extract data. As described above, the extraction is performed by using a statistical method such as a difference from the average face and principal component analysis as described above, or by template matching from gray images such as eyes, nose, and mouth. .
[0036]
The importance is set for each feature amount data. Here, the importance set in the feature amount data regarding eyes from the newly registered face data of “Mr. X” will be described as an example. The feature amount data relating to Mr. X's eyes and the feature amount data relating to the eyes of registrants other than Mr. X stored in the storage unit 30 are compared. Calculate the value.
[0037]
For example, a matching degree value obtained by comparing feature amount data relating to Mr. X's eye and feature amount data relating to Mr. A's eye, and a matching degree obtained by comparing feature amount data relating to Mr. X's eye and feature amount data relating to Mr. B's eye A comparison degree value is calculated individually by comparing this with feature amount data relating to the eyes of all registrants. Here, the comparison is made with all the registrants, but the comparison may be made with a certain number of registrants. Next, an average value for the matching degree value calculated for each registrant other than Mr. X is obtained, and this is the average of the feature amount data relating to Mr. X's eyes and the feature amount data relating to the eyes of registrants other than Mr. X. The collation degree value is used.
[0038]
FIG. 4 shows all the matching degree values calculated by comparing the registrants by the principal identification unit 26 with respect to the feature amount data relating to the eyes of the registrant stored in the storage unit 30. This is a graph of eye matching degree values. Here, the feature amount data relating to the eyes of all the registrants is compared. However, a certain number of registrants may be compared.
[0039]
In this graph, the average matching degree value between the feature data relating to the eyes of Mr. X and the feature data relating to the eyes of registrants other than Mr. X was calculated from the feature data relating to the eyes of all registrants. It turns out that it exists in the position away from the average collation degree value to the right side. In other words, the eyes of “Mr. X” are far from the feature amount data of the eyes of other registrants, and can be said to be feature amount data that can easily identify “Mr. X” at the time of collation determination. On the contrary, in this graph, the average matching degree value of the feature amount data regarding the eyes of Mr. X and the feature amount data regarding the eyes of the registrants other than Mr. X is obtained from the feature amount data regarding the eyes of all the registrants. If it is close to the calculated average collation degree value, it can be said that the feature amount data of “Mr. X” 's eyes is feature amount data that is not so important for identifying “Mr. X” at the time of collation determination.
[0040]
FIG. 5A shows the result of calculating the importance for each feature amount data. Average matching degree value (X) between the feature amount data relating to Mr. X's eye obtained above and the feature amount data relating to the eyes of registrants other than Mr. X _Eye (Average)) and the average matching degree value (all members) calculated from the feature data related to the eyes of all registrants _Eye (Average)) is calculated as the degree of importance. Specifically, all of the individual matching degree values calculated by comparing the feature amount data regarding all eyes of registrants and the average matching degree value calculated from the feature amount data regarding all eyes of registrants ( Everyone _Eye (Average)) and the standard deviation σ calculated from _Eye (Average) -X _Eye (Average))) is calculated as the absolute value of the value obtained by dividing. In this way, the importance level of the feature amount data related to Mr. X's eye is calculated, and the importance level is calculated for each feature amount data for the other feature amount data in the same manner. The calculated importance is assigned to each feature amount data, sorted in order of importance, and stored in the storage unit 30.
[0041]
FIG. 5B shows the feature quantity data in order of importance of “Mr. X” that stores and manages the importance assigned to each feature quantity in order of size. As a result, “Nose 3.5” and “XXX” with high importance of “Mr. X” are highly important, and mouth 1.2 and “△△△ 0.8” are low importance. It is done.
[0042]
The registration processing operation of the face collating apparatus 11 configured as described above will be described with reference to the flowchart of FIG.
[0043]
Now, when a person comes in front of the face collation device 11 and inputs the ID information of the registrant by inputting an ID code or an ID card (step n1),
The face image of the person to be registered is taken by the camera 12 (step n2).
[0044]
The photographed face image data is acquired by image capture, and stored in the image memory 23. After detecting the face area from the face image data by detecting the face area, if the position of the face is detected from the features such as eyes, nose, mouth, etc. and positioned (step n3),
Face feature data (see FIG. 3) is extracted from the face image (step n4).
[0045]
The extracted facial feature value data and the facial feature value data stored in advance in the storage unit 30 are compared by the person identifying unit 26 to individually calculate a matching degree value. Find the average value. For example, for the feature amount data for the eyes, the matching degree value is calculated individually by comparing with the feature amount data for the eyes of all registrants. Next, an average matching degree value is obtained from each registrant and the individually calculated matching degree value. This is performed for every feature data other than the eyes (step n5).
[0046]
The degree of separation between the obtained average matching degree value for each feature amount data and the average matching degree value for each feature amount data calculated from the feature amount data of all registrants is calculated as the importance degree for each feature amount data. To do. As described above, the importance is calculated for each feature amount data as an absolute value of a value obtained by dividing the deviation by the standard deviation σ. The calculated importance is assigned to each feature amount data and rearranged in order of importance (step n6).
[0047]
Importance is given to each feature quantity data, and the registration time is given to the face data rearranged in the order of importance (step n7), and is stored in the storage unit 30 together with the ID data. The registration process ends (step n8).
[0048]
Next, the collation processing operation of the face collation apparatus 11 will be described with reference to the flowchart of FIG.
[0049]
Now, when a person comes in front of the camera 12, the camera 12 captures the face of the person, acquires the captured face image data by image capture, and stores it in the image memory 23 (step n11).
[0050]
After searching the face area from the face image data and detecting the face area, the position of the face is detected from the features such as eyes, nose, mouth, etc. and positioned (step n12).
[0051]
Subsequently, face feature data is extracted from the face image data (step n13).
[0052]
The extracted feature value data of the face and the feature value data stored in advance in the storage unit 30 are compared and checked by the person identification unit 26 to obtain a matching degree value (step n14).
[0053]
If it is confirmed that the highest collation degree value is equal to or greater than the threshold based on the obtained collation degree value (step n15), the person who comes in front of the camera 12 is determined to be a registrant (step n16). ).
[0054]
At this time, the face data having the highest collation degree value, which is a source for discriminating whether or not the user is a registrant, is given the time when it was discriminated and re-registered in the storage unit 30 (step n17).
[0055]
Then, the result display 13 displays and outputs the fact that the user is a registrant, and the one verification process ends (step n18).
[0056]
On the other hand, in step n15 described above, if the highest collation degree value is equal to or less than the threshold value, it is not possible to identify the registrant.
[0057]
As described above, the registration time at the time of new registration assigned to the face data and the time at the time of determination are obtained from the face data assigned with the importance for each feature amount data stored in the storage unit. Next, a method for deleting data using the registration time at the time of new registration and the time when it has been determined given to the obtained face data will be described.
[0058]
For example, as shown in FIG. 8, a facial image composed of N feature data from feature 1 to feature N arranged in the order of high importance to low importance for registered data of a specific person registered in the storage unit 30. If the feature amount exists, data that has passed a certain elapsed time and data with low importance from the registration start time or the previous collation time are set as deletion targets.
[0059]
For example, if the Nth registered data with the lowest importance feature has passed 6 months, it will be deleted data, and the second least significant feature N-1 registration data will have passed one year. If it is, the third lowest importance feature N-2 is registered as the deletion target data if one year and six months have passed, and the fourth lowest importance feature N -The third registered data is carefully set and managed by comprehending the two deletion factors of elapsed time and importance so that it will be deleted if two years have passed.
[0060]
Next, the registration data deletion processing operation of the face matching device 11 will be described with reference to the flowchart of FIG.
[0061]
When deleting the feature amount data of the face data stored in the storage unit 30, first, the elapsed time calculation unit 28 calculates the elapsed time from the time given to the face data (step n21).
[0062]
At this time, if the elapsed time of the registered data is less than 6 months, for example, it is new face data and is not a deletion target (step n22).
[0063]
Similarly, when the elapsed time is checked and all data checks are completed, the deletion process ends (step n23).
[0064]
If there is face data having an elapsed time of 6 months or more to less than 1 year when the elapsed time of the registered data is checked (step n24),
Among the face data, the Nth feature data with the lowest importance is to be deleted, and the feature N (see FIG. 8) is deleted (step n25).
[0065]
Since the other feature data of the same person who was not the deletion target at this time is valid, the current time data is added to the face data including the valid feature data and re-registered in the storage unit 30 (step n26). ).
[0066]
Further, when the elapsed time is checked, if there is face data having an elapsed time of one year or more to less than one year and six months (step n27),
Among the face data, the Nth least important feature data and the N−1th feature data having the next lowest level are to be deleted, and the Nth feature and the N−1th feature are deleted (step n28).
[0067]
Furthermore, when the elapsed time is checked, if there is face data having an elapsed time of more than 1 year and 6 months to less than 2 years (step n29),
Among the face data, the Nth, N-1st, and N-2th feature data with the lowest importance are to be deleted, and the feature Nth, feature N-1th, and feature N-2nd Are deleted (step n30).
[0068]
Finally, when the elapsed time is checked, the data that exceeds the elapsed time of two years or more, the Nth, N-1st, and N-2th are the least important of the face data. , The N-3th feature data is a deletion target, and the feature Nth, the feature N-1th, the feature N-2th, and the feature N-3th are deleted (step n31).
[0069]
Even when one or more of these are deleted, other feature data of the same person who was not the deletion target is still valid, so the current time data is added to this valid data and re-registered in the storage unit 30. To do.
[0070]
In correspondence between the configuration of the present invention and the configuration of the above-described embodiment,
The storage means of the present invention corresponds to the storage unit 30 of the embodiment,
Similarly,
The discrimination means corresponds to the person identification unit 26,
The time registration means corresponds to the timer T,
The elapsed time calculation means corresponds to the elapsed time calculation unit 28,
The feature amount data deleting unit corresponds to the feature amount data deleting unit 27, but the present invention can be applied based on the technical idea shown in the claims and is limited to the configuration of the above-described embodiment. It is not something.
[0071]
【The invention's effect】
According to the present invention, since the deletion target data is clearly identified from the importance order feature quantity data stored and managed in order of the elapsed time and importance level of the registered data, the corresponding feature quantity data can be accurately deleted. Of the latest data, the most important data remains as registered data and can be updated for high collation management. In addition, it can be deleted in order of decreasing importance, and since it can eliminate the increase in data capacity, it does not take time to search for a face even when matching faces, and can be efficiently verified in a short time. Therefore, the present invention can be applied to a system having a small data capacity, and the cost is not increased because the data capacity is small.
[Brief description of the drawings]
FIG. 1 is a schematic configuration diagram showing a face matching device.
FIG. 2 is a control circuit block diagram of the face matching device.
FIG. 3 is an explanatory diagram showing facial feature data.
FIG. 4 is a chart showing a distribution state of eye matching degree values.
FIG. 5 is an explanatory diagram showing calculation results of importance for each feature amount data arranged in order of importance.
FIG. 6 is a flowchart showing a registration processing operation of the face matching device.
FIG. 7 is a flowchart showing a matching process operation of the face matching device.
FIG. 8 is an explanatory diagram showing a configuration of feature data and deletion points.
FIG. 9 is a flowchart showing a deletion processing operation of the face matching device.
[Explanation of symbols]
11. Face matching device
15 ... Control equipment
22 ... Control unit
26 ... Personal identification part
27. Feature data deletion unit
28: Elapsed time calculation unit
30 ... Storage unit
T ... Timer

Claims (3)

Storage means for associating face data composed of a plurality of feature amount data and registration time with all of the registrants, adding importance to the feature amount data, and storing the feature amount data;
Discrimination means for comparing the face data input for collation with the face data stored in the storage means to obtain a collation degree value, and discriminating whether the person is a registrant based on the obtained collation degree value;
When newly registering face data in the storage means, the face data is registered by giving a time at the time of registration, and when determining whether the determination means is a registrant, the memory used as a source for determination Time registration means for re-registering the face data that has been identified with the time when it was determined;
Based on the latest time given to the face data, elapsed time calculating means for calculating the elapsed time so far,
Feature value data deleting means for extracting face data whose elapsed time is equal to or longer than a predetermined time, and deleting the feature value data with low importance from the face data stored in the storage means;
A face matching device. The face collating apparatus according to claim 1, wherein the time registering unit gives a time when the original face data is determined when it is determined that the determining unit is a registrant. The importance is given to each feature amount data, and a registrant and others based on an average value of collation degree values obtained for each feature amount data by comparing feature amount data between the registrants. The face collation apparatus according to claim 1, wherein the face collation device is a separation degree data in which an average value of collation degree values obtained for each feature quantity data is compared by comparing feature quantity data with a registrant.

Images